Shazam, but for video.
You know a song from three seconds of it in a loud room. That is not magic, it is pattern matching, and it works because a song has a shape that survives being played badly. Video has one too. We learn the shape of yours, then we watch for it.
Below: the short version first, then the real mechanics for anyone who wants them, then the part where we tell you what it cannot do.
Four steps, no jargon.
- 01
We learn what your video looks and sounds like.
Not the file. The content. Every half second we take a reading of the picture and a reading of the sound, and turn both into numbers. A whole hour of video comes out as roughly two and a half megabytes of them.
- 02
We do the same thing to the big video sites, constantly.
Every video we can reach gets read the same way and added to the pile we search against. That pile is the entire product. Without it a set of numbers means nothing.
- 03
We look for stretches that line up.
Not one frame that looks similar. A run of seconds where your video and theirs agree, second after second, at a constant offset. Seven seconds of agreement is not a coincidence. One frame is.
- 04
You get the list.
Which site, which link, which seconds of your video appear at which seconds of theirs, and how sure we are. What you do next is yours to decide.
Two fingerprints, not one.
We run two completely separate ways of recognising your video, and we run both every time. They are not a backup for each other, they cover different ground. A screen recording usually only turns up in the sound results. A muted re-upload only turns up in the picture results. Run one of them and you throw away everything the other one could have seen.
Sound
We pick out the loudest points in the audio and record the pattern between them. This is the same trick a song recognition app uses to hear one track through a noisy bar.
- the picture being wrecked, blurred, cropped or filmed off a screen
- music laid over the top, which adds sound rather than removing it
- cuts and missing chunks in the middle
- being sped up or slowed down a little
- the audio being muted
- the audio being replaced with something else
- clips shorter than about five seconds
Picture
Every half second we turn the frame into a short string of numbers that stays roughly the same when the frame is re-encoded, recoloured or cropped. Similar frames get similar numbers.
- re-encoding, resizing and colour shifts
- cropping and letterboxing
- mirroring, because we search the flipped version too
- the sound being stripped out completely
- heavy re-encoding at very low quality
- screen recordings, which change the picture a lot
- clips shorter than about fifteen seconds
What actually counts as a match.
A single frame of your video looking like a single frame of theirs means nothing. At this scale every frame has thousands of accidental lookalikes. Similarity is not the signal.
The signal is agreement over time. We take a reading twice a second, so a ninety second clip is a hundred and eighty readings in a row. When all of them say the same thing, that your second 12 is their second 431, your second 12.5 is their second 431.5, and so on down the line, that is not a coincidence any more. It is the same footage.
Because we are matching a run rather than a whole file, gaps are free. Cut five seconds out of the middle and both halves still point at the same offset. Bury the clip an hour deep inside somebody else’s compilation and it still points at the same offset. Chop it into four pieces and reorder them and each piece reports itself separately.
Where it stops working.
Every tool in this space claims to find everything. None of them do, including this one. Here is the shape of what we cannot do, so you can judge the rest of the page against it.
About seven seconds, when the sound survives.
That is where the name comes from. Most re-uploads keep the original audio, because replacing it is work and it makes the clip worse.
About fifteen seconds, from the picture alone.
If the audio is muted or replaced, we need more of the picture to be confident. That is a real gap, and we would rather name it than hide it.
Under about five seconds, we cannot help.
There is not enough left to tell your video apart from every other video. No amount of engineering fixes that, and anyone claiming otherwise is selling something.
Still footage is harder than moving footage.
A locked-off shot where nothing changes gives us far less to line up than a scene that moves. The floors above assume ordinary footage.
These are the numbers we designed to, not measured results.
The honest version of this claim comes from cutting known clips at five, seven, ten, fifteen and thirty seconds and measuring where it breaks. We are building that test. When we have it, this page gets the real table.
We never claim to have searched everything.
Every result tells you which sites we searched and when. Nothing found on the sites we cover is useful information. It is not proof that nothing exists anywhere, and we will not let it look like one.
- We do not file anything for you. We find it, you decide what happens next.
- We never say a match is illegal. A copy that turns out to be licensed is a result you dismiss in two seconds, not a failure.
- What we give you is search output, useful for building your own case. It is not certified evidence and we will not dress it up as any.
- We do not touch content involving minors. That is a different job with different laws, and the right people already do it.
What actually leaves your computer.
Nothing you would recognise. The app reads your video where it sits, and sends back a list of numbers: the readings described above, and nothing else. No frames, no audio, no file. An hour of video goes in and about two and a half megabytes of arithmetic comes out.
Saying "we never store your content" is a promise you have to take on faith. Building it so the content never crosses the network at all is a property, and someone technical can watch their own network traffic while the app runs and see for themselves. If you are not that person, that is a fair thing to be unsure about, and the browser option exists for you: one video, searched, then deleted.
There is also a rule underneath all of it: whatever you send us to search for is never added to the pile we search against. Not a policy, a filter in the code that cannot be switched off.
Why we can give the app away.
Because the app is not the valuable part. A fingerprint on its own is inert. It only means something when it is matched against a pile of everything else, and that pile costs a great deal of continuous crawling to build and to keep current.
The index is the company. Which is also why we would rather tell you exactly which sites are in it than let you assume it covers the whole internet.
Things people ask about the tech.
Is this the same thing as reverse image search?
No. Reverse image search compares one still picture to other still pictures. We compare stretches of time. That is what lets us find a seven second slice of your video sitting in the middle of somebody else’s forty minute upload, which is the case that actually matters.
Is it the same as a file hash or checksum?
No, and this is the main reason older tools fail. A file hash changes completely if a single bit of the file changes, so re-encoding a video once defeats it. We do not look at the file at all. We look at what the video shows and what it sounds like.
Do you use face recognition?
No. Face recognition needs your face on file, and it stops working the moment the face is cropped out, blurred, turned away or simply not in the shot. We match the content instead, which is the one thing a leaker cannot remove and still have anything worth posting.
What if the video is mirrored or sped up?
Both are handled. We search the mirrored version of your video as well as the original, and we search at a spread of speeds from about 0.9x to 1.1x, which covers the range people actually use before it starts sounding wrong.
Can you find photos?
Not yet. Today this works on video and audio. If you have photos leaked, StopNCII is the better route right now and we will point you there rather than pretend otherwise.